Quantization of cepstral parameters for speech recognition over the World Wide Web

Abstract

We examine alternative architectures for a client-server model of speech-enabled applications over the World Wide Web. We compare a server-only processing model, where the client encodes and transmits the speech signal to the server, to a model where the recognition front end runs locally at the client and encodes and transmits the cepstral coefficients to the recognition server over the Internet. We follow a novel encoding paradigm, trying to maximize recognition performance instead of perceptual reproduction, and we find that by transmitting the cepstral coefficients we can achieve significantly higher recognition performance at a fraction of the bit rate required when encoding the speech signal directly. We find that the required bit rate to achieve the recognition performance of high-quality unquantized speech is just 2000 bits per second.2

Real Time Speech Based Integrated Development Environment for C Program

“…From the collected voice samples, it will identify the user moods. A client-server based speech recognition system is described in [8]. In this work, recognition is done at the server side.…”

Scalable distributed speech recognition using Gaussian mixture model-based block quantisation

“…We have chosen this method over the WER-based greedy algorithm of Digalakis et al (1999) because of its computational simplicity and this allows us to scale any bitrate with ease. Table 5 shows the average recognition accuracy of the non-uniform scalar quantiser.…”

“…Digalakis et al (1999) evaluated the use of uniform and nonuniform scalar quantisers as well as product code vector quantisers for compressing Mel frequencywarped cepstral coefficients (MFCCs) between 1.2 and 10.4 kbps. They concluded that split vector quantisers achieved word error rates (WER) similar to that of scalar quantisers while requiring less bits.…”

Efficient scalable encoding for distributed speech recognition

“…The effect of various speech coding techniques on speech recognition, including GSM (Digalakis et al, 1999;Srinivasamurthy et al, 2000;Kiss, 2000;Lilly and Paliwal, 1996;Srinivasamurthy et al, 2001b), G.723.1, G.727, G.728, G.729 (Turunen and Vlaj, 2001) (Lilly and Paliwal, 1996) and MELP (Srinivasamurthy et al, 2000;Srinivasamurthy et al, 2001b), has been previously evaluated by a number of researchers. In all cases, it was shown that speech coding significantly degrades speech recognition performance.…”

“…One such clientserver system, shown in Fig. 1, is distributed speech recognition (DSR) (Digalakis et al, 1999), where the client contains the feature extractor (the computation requirements of feature extraction are almost the same as those of a vocoder based speech encoder). Only the extracted features are encoded and transmitted to the server with the speech recognizer which operates on the decoded feature data.…”